Research map / Label-efficient and robust learning / Noisy labels
Graph convolutional networks for learning with few clean and many noisy labels
Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, Ondřej Chum, Cordelia Schmid
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GCN-based cleaning of noisy examples to class prototypes | Class prototypes · Low-shot ImageNet (1-shot) | top-5 accuracy: 45.3 -> 67.8 | Improved |
| add GCN-based cleaning of noisy examples to class prototypes | Class prototypes · Low-shot ImageNet (20-shot) | top-5 accuracy: 77.8 -> 78.2 | No clear effect |
| add noisy examples with unit relevance (beta=1) instead of cleaning (ablation) | Class prototypes · Low-shot ImageNet (20-shot) | top-5 accuracy: 77.8 -> 58.7 | Hurt |
| add graph propagation (GCN instead of MLP) (ablation) | relevance predictor · Low-shot ImageNet (1-shot) | top-5 accuracy: 63.6 -> 67.8 | Improved |
| add graph propagation (GCN instead of MLP) (ablation) | relevance predictor · Low-shot ImageNet (5-shot) | top-5 accuracy: 73.7 -> 73.9 | No clear effect |
| add learned GCN instead of label propagation (ablation) | relevance predictor · Low-shot ImageNet (1-shot) | top-5 accuracy: 62.6 -> 67.8 | Improved |
| replace class prototypes with cosine classifier learning (ablation) | GCN-cleaned data, ResNet-10 · Low-shot ImageNet (1-shot) | top-5 accuracy: 67.8 -> 73.2 | Improved |
| add deep network fine-tuning on top of cosine classifier (ablation) | GCN-cleaned data, ResNet-50 · Low-shot ImageNet (1-shot) | top-5 accuracy: 78.0 -> 80.2 | Improved |
| weight noisy examples by relevance instead of beta=1 (ablation) | Class prototypes · Low-shot ImageNet (1-shot) | top-5 accuracy: 56.1 -> 67.8 | Improved |
| add graph propagation (GCN instead of MLP) (ablation) | relevance predictor · Low-shot Places365 (1-shot) | top-5 accuracy: 46.9 -> 47.1 | No clear effect |
| add noisy examples with unit relevance (beta=1) (ablation) | Class prototypes · Low-shot Places365 (20-shot) | top-5 accuracy: 62.3 -> 50.8 | Hurt |
| add GCN cleaning of noisy data to prototypes | Class prototypes · Mini-ImageNet 5-way 5-shot | accuracy: 71.2 -> 74.7 | Improved |
Closest papers by meaning
- CleanNet: Transfer Learning for Scalable Image Classifier Training With Label Noise (CVPR 2018 · similarity 0.88)
- Learning Graph Convolutional Networks for Multi-Label Recognition and Applications (TPAMI 2021 · similarity 0.88)
- Simplifying Graph Convolutional Networks (ICML 2019 · similarity 0.88)
- FEW-SHOT LEARNING ON GRAPHS VIA SUPER-CLASSES BASED ON GRAPH SPECTRAL MEASURES (ICLR 2020 · similarity 0.87)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · similarity 0.87)
- Enhancing the Influence of Labels on Unlabeled Nodes in Graph Convolutional Networks (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
